2026-07-31·Daily digest

🔬 Research Pulse

Daily Digest

July 31, 2026

🤖 AI

🧠 LLMs

1. AISPA: User-Centric System Prompt Auditing for Large Language Model Applications

Authors: Xiangning Lin, Shenzhe Zhu, Shu Yang... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can we systematically audit undisclosed system prompts in commercial LLM applications to close the trust and accountability gap between developers and end users?

Summary: AISPA introduces a user-centric auditing framework that evaluates system prompts along 8 user-relevant dimensions, classifying each instruction as protective or problematic. Applying it to 3,249 instructions from 88 commercial AI products reveals broad but shallow adoption of user protections and pervasive coexistence with problematic instructions, arguing for transparency and independent oversight.

Key Results: Analyzed 3,249 instructions from system prompts across 88 commercial AI products using the 8-dimension AISPA taxonomy. Found: (1) 98.9% of products contain at least one protective instruction, but only 24% cover all 8 dimensions; (2) instruction counts per product vary from <5 to >60 across organizations; (3) system prompts have grown longer and more protective over time; (4) ~40% of products contain at least one problematic instruction working against user interests, often coexisting with protective ones.

Key Findings:

98.9% of products have at least one protective instruction, but only 24% cover all 8 AISPA dimensions — protection is wide but shallow
Enormous variance across developers (from <5 to >60 protective instructions per product) indicates no industry standard
~40% of products contain instructions that actively work against user interests, frequently coexisting with protective ones in the same prompt
System prompts have grown longer and more user-protective over time, suggesting emerging norms

Technical Novelty: First user-centric (rather than developer- or safety-centric) auditing framework for system prompts, with an 8-dimension taxonomy classifying individual instructions as protective vs. problematic — enabling instruction-level rather than prompt-level analysis.

What's New: Prior work focused on prompt extraction, jailbreaking, or safety-from-developer perspective. AISPA reframes system prompts as a governance/accountability artifact evaluated from the user's interest standpoint, with an operationalized 8-dimension taxonomy applied at empirical scale.

Extension Opportunities:

Build an automated real-time monitoring tool that extracts and scores system prompts against the AISPA 8-dimension taxonomy for browser-accessible AI products
Extend the taxonomy to cover agentic systems with tool-use and multi-step reasoning, where system prompts include tool definitions and safety constraints
Develop adversarial techniques or user-facing 'prompt reveal' interfaces so regulators can independently verify compliance without vendor cooperation

Replicability: Abstract does not mention released code or dataset. Reproduction would require re-collecting 88 products' system prompts (via jailbreaking/extraction techniques) and manual or LLM-assisted annotation of ~3,249 instructions — moderate labor, low compute.

Research Gaps:

No mechanism for continuous or independent verification as prompts change post-audit
Taxonomy grounded in current chat products may not generalize to agentic/multi-modal systems or embedded enterprise deployments

2. OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models

Authors: Qiushi Sun, Kanzhi Cheng, Yian Wang... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: Are VLM judges reliable enough to verify computer-use agent (CUA) trajectories at scale for evaluation, data curation, and RL reward signals?

Summary: OSReward is a human-annotated benchmark for evaluating VLM judges of computer-use agent trajectories across platforms. It reveals that frontier VLM judges share a systematic leniency bias, and introduces OS-Shepherd (9B/35B) — open reward models trained on a released 100K corpus that match commercial judges at 30-60% lower cost.

Key Results: Introduced OSReward benchmark (plus OSReward-Hard and OSReward-Multi) with multi-stage human-annotated ground truth across diverse agent backbones and platforms. Demonstrated that state-of-the-art VLM judges systematically exhibit leniency bias, mislabeling failed runs as successes. Trained OS-Shepherd (9B and 35B) on the released OS-Shepherd-100K corpus, matching commercial judge quality at 30-60% lower cost.

Key Findings:

All evaluated VLM judges, including SOTA, fall short of an ideal judge and share a systematic leniency bias mislabeling failures as successes
Reliable proprietary judges are prohibitively expensive at scale, while affordable open models trail significantly in accuracy
OS-Shepherd (9B/35B) trained on OS-Shepherd-100K matches commercial judge reliability at 30-60% lower cost, providing a viable open alternative

Technical Novelty: First standardized, human-verified cross-platform benchmark specifically targeting VLM-as-judge reliability for CUA trajectories, coupled with an open reasoning-annotated judgment corpus (OS-Shepherd-100K) and open reward models that close the cost/quality gap with proprietary judges.

What's New: Prior work assumed VLM-as-judge reliability without systematic auditing; this paper is the first to construct a rigorous cross-platform, human-verified benchmark for CUA trajectory judgment and to release open reward models plus training data addressing the identified gap.

Extension Opportunities:

Use OS-Shepherd as a reward model in online RL loops for CUA training and measure downstream task success gains vs. commercial-judge-based RL
Extend the benchmark to mobile OS and browser-only agents, or to long-horizon multi-app trajectories where leniency bias likely compounds
Build a debiasing layer that calibrates VLM judges against OSReward-Hard to correct the identified leniency bias without retraining

Replicability: Code, benchmark, dataset, and 9B/35B checkpoints are publicly released at os-copilot.github.io/OSReward-Home. Reproducing evaluation is lightweight (inference only); training the 35B reward model requires substantial multi-GPU compute (likely 8+ high-memory GPUs).

Research Gaps:

Root causes of the leniency bias (training data artifacts vs. visual grounding limits) are not fully diagnosed
Impact of using OS-Shepherd as an RL reward signal on downstream CUA performance is not yet quantified end-to-end

3. Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments

Authors: Haomin Qi, Xingliang Wang, Xuanqi Gao... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can we scale the supply of executable, verified coding-agent tasks (with realistic states, specs, tools, and verification) without repeatedly paying the cost of building fresh environments for each new task?

Summary: Change2Task converts merged pull requests into executable, verified coding-agent tasks placed on healthy modern revisions of the same repo, using patch reversal, code mapping, or agent reconstruction. It reports 79.6% verified construction success on 1,130 candidates across five task families, 29.2% more verified tasks than a PR baseline, and 10.8% lower pipeline cost via base reuse.

Key Results: Starting from 1,130 eligible source changes, Change2Task achieved 79.6% verified task construction success across five task families (Bug Fix, Feature Addition, Test Generation, API Migration, Security Repair). On a matched candidate set it recovered 29.2% more verified tasks than a PR-based construction baseline. Historical vs reconstructed cases reached up to 98.0% matched outcome agreement under agent evaluation, and reusing modern bases reduced full-pipeline measured expenditure by 10.8%.

Key Findings:

Reconstructing tasks on modern bases (rather than historical commits) yields 29.2% more verified tasks than a PR-based baseline on the same candidate set.
Reconstructed task outcomes agree with historical outcomes up to 98.0% under agent evaluation, suggesting fidelity is preserved.
Sharing a maintained modern base across many tasks cuts full-pipeline expenditure by 10.8%, addressing a real scaling bottleneck.

Technical Novelty: Rather than freezing a task at its historical commit (SWE-bench style), Change2Task ports historical PR evidence forward onto a healthy modern revision using three reconstruction strategies — Patch Reversal, Code Mapping, and Agent Reconstruction — and validates a full base→task→restored lifecycle. This amortizes environment setup across many tasks derived from one maintained base.

What's New: Prior benchmarks (SWE-bench, SWE-Gym) tie each task to its original historical commit and rebuild environments per task. Change2Task decouples the task specification from its commit and re-hosts it on a maintained base, and it generalizes beyond bug fixes to five task families with a unified verification lifecycle.

Extension Opportunities:

Add more task families beyond the five studied (e.g., performance optimization, dependency upgrade, refactoring, concurrency bug fixes) using the same three reconstruction primitives.
Apply Change2Task to non-Python ecosystems (JS/TS, Rust, Go) where PR-to-task pipelines like SWE-bench are underrepresented, and publish a multi-language benchmark.
Use the generated verified tasks as an RL training signal for coding agents, comparing agent skill gains against SWE-Gym or SWE-bench-trained baselines.

Replicability: The abstract does not mention released code, data, or a public artifact. Reproduction would require a large PR corpus mining pipeline, sandboxed execution environments for many repos, and an LLM agent capable of reconstruction and evaluation — likely tens of thousands of container-executions plus significant LLM inference budget.

Research Gaps:

No evidence yet on cross-language generality — the abstract does not specify which ecosystems were tested.
The 98% outcome-agreement figure needs finer analysis of where reconstruction diverges from historical behavior and whether that biases agent evaluations.

🦾 ROBOTICS

1. DexDirect: Direct Kinesthetic Arm Guidance for Efficient Dexterous Demonstration Collection

Authors: Beom Jun Kim, Shiu-Jen Wang, Jonathan Liu... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can dexterous manipulation demonstrations be collected efficiently at scale without expensive hardware or high cognitive load on operators, addressing the tradeoff between fidelity and setup cost in existing teleoperation interfaces?

Summary: DexDirect is a hybrid teleoperation interface where operators physically drag a gravity-compensated robot arm by a handle while a webcam retargets their other hand's pose onto a dexterous 13-DoF end-effector. It substantially outperforms pure-vision and pose-tracking baselines in demonstration throughput and operator workload, and produces data sufficient to train a diffusion policy to 90% success on pick-and-place.

Key Results: DexDirect collected 17.2x more successful demonstrations than AnyTeleop (vision-based) and 3.2x more than TeleDex (pose-tracking) baselines. A diffusion policy trained on DexDirect data achieved 90% success on a cube pick-and-place task. NASA-TLX evaluations showed significant reductions in mental demand, effort, and frustration, at the cost of increased physical demand.

Key Findings:

Direct kinesthetic arm guidance yields 17.2x and 3.2x more successful demos than AnyTeleop and TeleDex respectively
Splitting control across modalities (physical arm + visual hand) reduces cognitive load — mental demand, effort, and frustration all drop, though physical demand rises
Data quality is high enough that a diffusion policy trained on these demos reaches 90% success on a cube pick-and-place task

Technical Novelty: Novel hybrid interface combining direct physical dragging of a gravity-compensated 6-DoF arm (kinesthetic) with webcam-based vision retargeting of the operator's other hand onto a 16-joint 13-DoF dexterous hand — decoupling arm and hand control modalities to leverage the strengths of each.

What's New: Prior work forces a choice between high-fidelity but expensive setups (exoskeletons, VR with haptics) and cheap but imprecise vision-only teleoperation. DexDirect is novel in hybridizing modalities — kinesthetic for the arm (precise, low-cognitive-load) and vision for the hand (cheap, no wearable) — rather than picking one paradigm for the whole system.

Extension Opportunities:

Extend the single-arm setup to bimanual kinesthetic guidance for two-handed manipulation tasks like assembly or cloth folding
Replace the single webcam hand retargeting with multi-view or depth-based retargeting to handle occlusions during in-hand manipulation
Scale up data collection across diverse tasks and train a multi-task foundation policy, benchmarking against demonstrations collected via VR or exoskeleton interfaces

Replicability: The abstract does not mention code or dataset release. Reproduction would require a gravity-compensated 6-DoF robot arm with force-torque handle, a 16-joint dexterous hand (e.g., Allegro/LEAP-class), a webcam, and compute for diffusion policy training (typically a single modern GPU for the scale implied).

Research Gaps:

No evaluation on precision-critical or contact-rich tasks beyond cube pick-and-place; unclear how the interface scales to insertion, tool use, or deformable objects
Increased physical demand may limit long-session data collection throughput; ergonomic tradeoffs and operator fatigue over hours/days are not characterized

2. RedFlow: Redirect Failure into Action-Level Corrections for Flow-matching VLA Policy

Authors: Zhengyang Yan, Junhao Li, Fangqi Zhu... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can offline RL leverage failure data at a fine-grained action level (rather than trajectory level) to correct compounding distribution-shift errors in flow-matching VLA policies?

Summary: RedFlow is an offline RL framework for flow-matching VLA policies that converts failed rollouts into dense action-level corrective supervision by retrieving successful alternatives from similar contexts. It jointly reinforces good actions, suppresses bad ones, and redirects recoverable failures — outperforming offline baselines and matching on-policy RL with ~10x less data.

Key Results: On LIBERO benchmark and 3 real-world manipulation tasks, RedFlow raised real-world success from 56.7% to 74.7% (+18 pts), beat SOTA offline RL baselines, and matched on-policy methods (PPO, GRPO, DDPO) with ~10x fewer training samples.

Key Findings:

Action-level (rather than trajectory-level) exploitation of failure data substantially improves offline RL for VLA policies.
Retrieving successful alternatives from contextually similar states provides usable corrective targets even from mixed-quality data.
Real-world success rate improved 56.7% → 74.7%, matching on-policy PPO/GRPO/DDPO at roughly an order of magnitude lower sample cost.

Technical Novelty: Two novel components: (1) Context-Aware Corrective Matching that identifies specific failure-inducing actions and retrieves successful alternatives from contextually similar states, and (2) Adaptive Redirection Objective that unifies three signals (reinforce success, suppress failure, redirect recoverable failures) into flow-matching supervision — versus prior offline RL that either discards failure data or weights it only at trajectory level.

What's New: First offline RL framework to convert failure experiences into fine-grained action-level corrective targets specifically tailored to flow-matching VLA policies, unifying reinforcement, suppression, and redirection in a single objective.

Extension Opportunities:

Apply the Context-Aware Corrective Matching mechanism to diffusion-based (non-flow-matching) VLA policies like Octo or RDT to test generalization across generative policy classes.
Extend the Adaptive Redirection Objective to online/interactive settings where corrective targets are retrieved from a continually growing rollout buffer.
Use RedFlow's action-level failure identification as a signal for active data collection — targeting human demonstrations only at contexts where retrieval finds no successful alternative.

Replicability: Abstract does not mention code release. LIBERO is a public benchmark; reproduction would require a flow-matching VLA backbone (e.g., pi0-style), a modest offline RL compute budget (paper claims ~10x fewer samples than PPO/GRPO/DDPO), and real robot hardware for the 3 manipulation tasks.

Research Gaps:

Prior offline RL for VLAs either discards failure trajectories or applies coarse trajectory-level weighting, wasting signal.
Flow-matching policies lack tailored RL fine-tuning methods that exploit their per-step generative structure.

3. X-NavDP: Generalizing Navigation Diffusion Policy to Novel Behavior and Embodiments with Group Q-score Reweighted Matching

Authors: Tianyu Yang, Yiming Zeng, Wenzhe Cai... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can navigation diffusion policies generalize beyond oracle-generated expert demonstrations to novel embodiments and challenging local-reactive scenarios (dead ends, long obstacles), given that RL post-training of diffusion policies has historically yielded only marginal gains due to intractable likelihoods and inefficient exploration?

Summary: X-NavDP proposes GQRM, a diffusion-policy RL post-training framework that overcomes the instability of prior methods via group Q-score reweighted score matching and prior-preserving perturbation-based exploration. Applied to cross-embodiment visual navigation, it lifts success rates by ~23 points in sim and 55 points on real-world hard cases over pretrained baselines.

Key Results: Introduces GQRM (Group Q-score Reweighted Matching) for diffusion RL post-training. The resulting X-NavDP policy improves cross-embodiment visual navigation success rate from 61.20% to 84.28% in simulation (+23.08 pts) and from 10% to 65% in real-world hard cases (+55 pts), demonstrated across heterogeneous embodiments via distributed online RL.

Key Findings:

Pretraining on oracle demonstrations fundamentally caps diffusion navigation policies on reactive behaviors like dead-end escape and long-obstacle detour
Reweighted score matching using per-trajectory group-normalized Q-scores is a stable substitute for intractable diffusion policy gradients
Distributed online RL across heterogeneous embodiments produces a single policy that generalizes cross-embodiment, achieving 84.28% sim / 65% real success

Technical Novelty: Two coupled contributions: (1) self-bootstrapped exploration with behavior perturbation that preserves the pretrained diffusion prior (avoiding mode collapse common in RL fine-tuning), and (2) group Q-score normalization computing per-trajectory values at each state for reweighted score matching — sidestepping the intractable diffusion likelihood that destabilizes prior policy-gradient approaches (e.g., DDPO, DPPO).

What's New: First diffusion-RL framework to yield substantial (not marginal) gains for navigation by combining behavior-perturbed self-bootstrapped exploration with group-normalized reweighted score matching, trained cross-embodiment online rather than per-robot.

Extension Opportunities:

Apply GQRM's group Q-score reweighted matching to other diffusion-based robot policies (manipulation, humanoid locomotion) where oracle demos also limit generalization
Extend the self-bootstrapped exploration with behavior perturbation to language-conditioned navigation, allowing RL fine-tuning of VLA navigation models
Investigate sim-to-real gap reduction by combining GQRM with domain randomization or real-world offline RL data to further close the 84% sim vs 65% real gap

Replicability: Code and model publicly available at https://yty-sky.github.io/x-navdp-project-page. Reproduction requires distributed online RL infrastructure across heterogeneous simulated embodiments — likely multi-GPU (8+ A100/H100 class) for diffusion policy training plus a photorealistic navigation simulator (e.g., Habitat/Isaac). Real-world validation needs multiple robot platforms.

Research Gaps:

Real-world performance (65%) still trails simulation (84%), indicating unresolved sim-to-real transfer challenges for diffusion navigation policies
The framework is demonstrated only on visual navigation; whether GQRM scales to longer-horizon, semantic, or language-conditioned tasks remains open

💻 COMPUTE

1. A Cloud Continuum Research Infrastructure for Distributed CPS Experimentation

Authors: Fabio Orazio Mirto, Giuseppe Tricomi, Luca D'Agati... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can researchers systematically experiment with Cyber-Physical System (CPS) workflows across heterogeneous Edge-Fog-Cloud continuum infrastructures while preserving reproducibility, observability, and control over distributed deployments — beyond single-domain prototypes?

Summary: The paper introduces a two-level reference architecture for Cloud Continuum experimentation on top of the SLICES Blueprint, cleanly separating infrastructure management from CPS application workflows organized in an Edge-Fog-Cloud pattern. It is validated through two workloads — Renewable Energy Community management and AirWatch air-quality monitoring — across 40 experimental runs comparing virtualized and physical edge deployments over a geographically distributed testbed.

Key Results: The authors validated a two-level reference architecture built on the SLICES Cloud Continuum Blueprint through two use cases: (1) Renewable Energy Community management with distributed Digital Twin coordination and time-window energy control, and (2) AirWatch, an anomaly-detection and low-latency alerting pipeline. Evaluation consisted of a systematic campaign of 40 runs comparing virtualized vs. physical edge deployments across a geographically distributed infrastructure. Specific latency/throughput numbers are not disclosed in the abstract.

Key Findings:

A generic (non-domain-specific) two-level architecture can host heterogeneous CPS applications while preserving reproducibility and observability.
The Edge-Fog-Cloud partitioning (low-latency sensing/actuation → near-source coordination/stream processing → global analytics/optimization) is expressive enough to cover both Digital-Twin coordination and anomaly-detection pipelines.
Systematic 40-run comparisons between virtualized and physical edge deployments over a distributed infrastructure are feasible on top of the SLICES Blueprint.

Technical Novelty: The clean separation between a research-infrastructure layer (exposing/managing distributed heterogeneous resources) and an application layer (organizing CPS workflows via Edge-Fog-Cloud partitioning), with placement, timing, and data provenance elevated to first-class experimental concerns — rather than a domain-locked prototype. Built specifically on the SLICES Cloud Continuum Blueprint for multi-application programmability.

What's New: Prior continuum experimentation work has largely focused on single-domain prototypes or on infrastructure abstractions alone. This paper couples the SLICES Blueprint with a reusable, multi-application experimental layer that treats placement, timing, and data provenance as first-class variables under researcher control.

Extension Opportunities:

Add a third use case in a different domain (e.g., industrial IoT, smart mobility, or precision agriculture) to further generalize the architecture and stress-test placement policies under different timing constraints.
Integrate an automated placement optimizer (RL- or ILP-based) that treats placement, timing, and provenance as jointly optimized decision variables rather than fixed experimental configurations.
Extend the observability layer with fine-grained energy/carbon accounting per continuum tier to enable sustainability-aware experimental comparisons of virtualized vs. physical edge deployments.

Replicability: The abstract does not mention released code, datasets, or an artifact. Reproduction would require access to a SLICES-compatible geographically distributed testbed with Edge, Fog, Cloud, and HPC tiers — non-trivial infrastructure typically only accessible via federated European research testbeds.

Research Gaps:

Concrete quantitative performance results (latency, throughput, jitter, resource utilization) for the 40-run campaign are not surfaced in the abstract.
No mention of automated or policy-driven placement decisions — placement appears to remain a manual experimental parameter rather than a solved optimization problem.

2. A Taxonomy of Performance Metrics for the Distributed Computing Continuum

Authors: Praveen Kumar Donta, Boris Sedlak, Alfreds Lapkovskis... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can performance metrics for Distributed Computing Continuum Systems (DCCS) be organized into a unified, cross-layer taxonomy, given that existing evaluation practices focus on isolated dimensions (compute, network, energy, or application quality) and fail to capture holistic behavior across the edge-to-cloud continuum?

Summary: The paper introduces a structured taxonomy of performance metrics for Distributed Computing Continuum Systems that spans computing-, network-, and application-level dimensions and adds six emerging cross-cutting dimensions (sustainability, observability, adaptability, data locality, migration awareness, continuum fragmentation). Each metric is given a mathematical formulation and characterized by acquisition scope, phase, and method, filling the gap left by prior evaluations that only examined isolated layers.

Key Results: This is a taxonomy/survey paper rather than an empirical study — it does not report benchmark numbers or datasets. Its concrete contribution is a structured classification organizing DCCS metrics into three primary categories (computing-level, network-level, application/user-level) plus six emerging dimensions (sustainability, observability, adaptability, data locality, migration awareness, continuum fragmentation), with mathematical formulations for each and a three-axis acquisition framework (scope: single-node/multi-node/system-wide; phase: operational vs. experimental; measurement method).

Key Findings:

Existing DCCS evaluations are fragmented across compute, network, energy, and QoS silos and cannot capture cross-layer behavior
Six emerging dimensions — sustainability, observability, adaptability, data locality, migration awareness, and continuum fragmentation — are essential but underrepresented in current metric catalogs
Metrics differ meaningfully by acquisition scope (single node vs. multi-node vs. system-wide) and by whether they suit operational monitoring or experimental evaluation, and this distinction should drive methodology choice

Technical Novelty: Prior surveys treat compute, network, and energy metrics in silos; this work's novelty is (1) unifying them under a single DCCS-oriented taxonomy, (2) elevating six emerging cross-cutting dimensions (notably migration awareness and continuum fragmentation) that don't appear in traditional distributed-systems metric catalogs, and (3) formalizing metric acquisition along scope/phase/method axes so practitioners can decide what is measurable where.

What's New: First taxonomy explicitly designed for the continuum (rather than cloud, edge, or HPC in isolation) that treats migration awareness and continuum fragmentation as first-class metric families and pairs each metric with a formal acquisition-requirement classification.

Extension Opportunities:

Build an open-source reference benchmarking harness that instruments the taxonomy's metrics against real DCCS workloads (e.g., federated inference across K3s edge clusters + cloud GPUs) to produce the empirical baseline the paper itself lacks
Operationalize the 'migration awareness' and 'continuum fragmentation' dimensions as first-class signals in a Kubernetes/KubeEdge scheduler, using them as inputs to placement decisions for AI inference workloads
Extend the taxonomy with LLM-serving-specific metrics (tokens/sec/watt, KV-cache locality, speculative-decoding hit rates across edge/cloud tiers) since the abstract flags AI as a driver but the categories remain generic

Replicability: As a taxonomy paper, replicability is not applicable in the experimental sense — no code, datasets, or compute requirements are indicated in the abstract. Reuse would consist of adopting the mathematical formulations and acquisition-requirement framework in one's own evaluation methodology.

Research Gaps:

No empirical validation — the taxonomy is not exercised against real workloads to show which metrics actually discriminate between DCCS designs
AI/LLM-specific serving metrics (throughput-per-watt, KV-cache behavior, disaggregated prefill/decode) are motivated but not concretely developed within the categories

3. NELSSA: A GPU-PNM Heterogeneous System for Mixed-Length LLM Serving via Length-based Request Placement

Authors: Sookyung Choi, Seungyong Lee, Kangkyu Park... Published: 2026-07-29 | Citations: 0 arXiv | PDF

Research Question: How can LLM serving systems efficiently handle mixed-length workloads (from hundreds to hundreds of thousands of tokens) that expose inefficiencies in GPU-centric architectures whose throughput depends on large, memory-constrained batches?

Summary: NELSSA is a heterogeneous GPU-PNM LLM serving system that routes short-context requests to GPUs and long-context requests to Processing-near-Memory accelerators via CXL disaggregation, with runtime migration to handle context growth. The prototype achieves up to 5.5x decode throughput and 15x P99 latency improvements over GPU-only baselines on mixed-length workloads.

Key Results: NELSSA prototype demonstrates up to 5.5x improvement in decode throughput (tokens/sec) and up to 15x reduction in P99 latency on mixed-length LLM workloads compared to GPU-only baselines. Built as end-to-end system with device-level sparse attention on PNM, GPU decode kernels, host-side runtime, and CXL-based RPC/RDMA infrastructure.

Key Findings:

GPU-centric serving is fundamentally inefficient for mixed-length workloads because throughput depends on large memory-constrained batches that long contexts break
Length-based request placement combined with cross-tier runtime migration is a viable scheduling primitive for heterogeneous serving
CXL-based disaggregation with RPC/RDMA can practically bridge GPU and PNM tiers for production LLM inference

Technical Novelty: First end-to-end prototype integrating real (not simulated) PNM accelerators with GPUs via CXL for LLM serving, combined with length-based request routing and runtime migration that avoids recomputation when short contexts grow into long ones. Prior PNM/PIM LLM work has largely been simulation-based or single-tier.

What's New: Moves beyond simulation-based PIM/PNM LLM research to a working end-to-end prototype, and introduces length-based tier placement with non-recomputing migration — a scheduling approach specifically designed for the bimodal nature of modern agentic LLM traffic.

Extension Opportunities:

Extend length-based placement to a multi-tier hierarchy (GPU/PNM/CPU-DRAM/SSD) with predictive migration based on request growth patterns
Apply the GPU-PNM heterogeneous architecture to multi-modal serving where vision/audio tokens have different memory-bandwidth profiles than text
Develop cost-aware autoscaling policies that provision GPU:PNM ratios dynamically based on observed workload length distributions

Replicability: No explicit mention of code/data release in abstract. Reproduction would require access to real PNM accelerator hardware (likely Samsung/SK Hynix HBM-PIM or AXDIMM class devices), CXL-enabled infrastructure with RDMA fabric, and GPUs — a highly specialized hardware stack unavailable to most researchers.

Research Gaps:

No discussion of prefill-phase handling or how placement decisions interact with speculative decoding and disaggregated prefill/decode architectures
Unclear how the system behaves under adversarial or highly skewed length distributions, or when migration itself becomes a bottleneck

⚡ ENERGY

1. Fast and Accurate Foundation Models for Equivariant Machine-Learned Interatomic Potentials

Authors: Seán R. Kavanagh, Chuin Wei Tan, Menghang Wang... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can equivariant machine-learned interatomic potentials (MLIPs) simultaneously achieve high inference/training speed AND high accuracy when scaled to extremely large foundation-model datasets, where the data-efficiency advantage of equivariance matters less?

Summary: The paper introduces a family of accelerated foundation potentials in the NequIP and Allegro equivariant MLIP frameworks that achieve state-of-the-art inference speed and scalability while remaining competitive on materials-discovery, thermal-conductivity, and near-equilibrium property benchmarks. It also argues that further accuracy gains in universal potentials will come from broader dataset diversity and improved transition-metal PES coverage, not architectural tweaks alone.

Key Results: The authors present a family of foundation potentials built on the NequIP and Allegro equivariant architectures that deliver leading inference speeds, strong scalability, and competitive accuracy across community benchmarks spanning materials discovery, thermal conductivity prediction, and near-equilibrium mechanical/thermodynamic properties. Accelerations in the NequIP infrastructure dramatically reduce the cost of training high-accuracy foundation models on ultra-large datasets. They also identify that accuracy gains for materials discovery come primarily from dataset diversity and better/more consistent descriptions of transition-metal-compound potential energy surfaces (specific numeric benchmark values are not quoted in the abstract).

Key Findings:

Equivariant MLIPs can be engineered for leading inference speed at foundation-model scale without giving up accuracy, refuting the common assumption that equivariance is inherently too slow for large-data regimes.
New NequIP/Allegro training accelerations sharply reduce the cost of building high-accuracy foundation models on ultra-large datasets.
Materials-discovery accuracy is bottlenecked more by dataset diversity and inconsistent transition-metal-compound energy surfaces than by model architecture.

Technical Novelty: Architectural and infrastructure-level accelerations to NequIP/Allegro that break the conventional speed-vs-accuracy trade-off for equivariant models at foundation-model scale, plus the empirical demonstration that at ultra-large data regimes the data-efficiency premium of equivariance can be traded for throughput without sacrificing accuracy.

What's New: Prior equivariant foundation MLIPs (NequIP, Allegro, MACE) were seen as accurate but slower than invariant/scalar alternatives; this work shows equivariant models can match or beat them on throughput at scale while retaining accuracy, and explicitly redirects the field's improvement agenda toward data curation, especially for transition metals.

Extension Opportunities:

Fine-tune the released NequIP/Allegro foundation potentials on domain-specific datasets (e.g., catalysts, battery electrolytes, MOFs) and quantify the speed/accuracy Pareto vs. MACE, SevenNet, or Orb baselines.
Build a curated 'transition-metal-hard-cases' benchmark and retraining set targeting the inconsistent DFT-level PES descriptions the authors flag, then measure how much this closes the materials-discovery accuracy gap.
Integrate the accelerated NequIP kernels into a large-scale MD workflow (LAMMPS/ASE) and benchmark million-atom simulations for thermal conductivity or mechanical response against classical potentials and other MLIP foundation models.

Replicability: The paper implies release via the NequIP infrastructure (an established open-source codebase), suggesting model weights and training code will be available. Full reproduction of foundation-model training would require substantial GPU compute (multi-GPU, likely hundreds to thousands of GPU-hours on ultra-large DFT datasets such as OMat24/MPtrj-scale), though inference and fine-tuning should be feasible on a single modern GPU.

Research Gaps:

Lack of consistent, high-quality DFT reference data for transition-metal compounds across foundation-model training sets.
Absence of universal benchmarks that jointly measure inference throughput, training cost, and accuracy on downstream MD tasks rather than energy/force RMSE alone.

2. Transport Evidence of Magnetic Polarization in the Altermagnetic Candidate MnTe

Authors: Younes Ghorbani, Nayana Devaraj, Joshua Maile... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can a finite net magnetic polarization be engineered in α-MnTe—nominally a compensated antiferromagnet and altermagnetic candidate—through thin-film epitaxy, and what transport signatures reveal it?

Summary: The authors grow epitaxial α-MnTe on InP(111) by MBE and observe hysteretic butterfly longitudinal and nonlinear transverse magnetoresistance, transport signatures of a finite net magnetization in a material normally considered a compensated antiferromagnet. DFT attributes the emergent polarization to interface-induced symmetry breaking and point defects, establishing an epitaxial route to engineer magnetic response in altermagnetic candidates.

Key Results: Epitaxial α-MnTe films were grown on InP(111) by MBE. Low-temperature magneto-transport shows (1) a hysteretic butterfly-shaped longitudinal magnetoresistance and (2) a nonlinear transverse magnetoresistance, both consistent with a finite net magnetization. DFT calculations identify interface-induced symmetry breaking and point defects as microscopic pathways producing uncompensated moments in thin-film geometries. No specific field ranges, temperatures, or MR percentages are quoted in the abstract.

Key Findings:

Successful MBE growth of epitaxial α-MnTe on InP(111).
Hysteretic butterfly longitudinal magnetoresistance and nonlinear transverse magnetoresistance indicate net magnetic polarization.
DFT identifies interface symmetry breaking and point defects as mechanisms yielding uncompensated moments in thin films.

Technical Novelty: First transport-level demonstration that thin-film α-MnTe (an altermagnet candidate) hosts a measurable net magnetization via butterfly longitudinal + nonlinear transverse MR on InP(111), paired with DFT identifying interface and defect mechanisms rather than treating MnTe as a strictly compensated antiferromagnet.

What's New: Combines a new substrate/film system (α-MnTe on InP(111)) with transport evidence of emergent ferromagnetic-like response in an altermagnet candidate and a DFT-grounded microscopic explanation—prior work largely treated MnTe as fully compensated.

Extension Opportunities:

Systematically vary film thickness and substrate (e.g., GaAs(111), Al2O3) to disentangle interface symmetry breaking from bulk defect contributions to the observed hysteresis.
Introduce controlled Mn/Te vacancies via growth stoichiometry or post-annealing and correlate defect density (measured by STEM/APT) with transverse MR magnitude to validate the DFT defect pathway.
Fabricate MnTe-based spin-valve or tunnel junction devices to test whether the emergent polarization enables electrical readout of altermagnetic order for spintronic memory.

Replicability: No code/data availability mentioned. Reproduction requires an MBE system with InP(111) substrates, cryogenic magneto-transport (dilution/PPMS-class), and DFT resources (VASP/QE-scale, ~modest HPC for supercells with defects and interfaces).

Research Gaps:

No direct magnetometry (SQUID/XMCD) quantifying the magnetization magnitude to corroborate transport inferences.
Relative contributions of interface vs. defect mechanisms are computationally suggested but not experimentally disentangled.

3. Quasiparticle phono-conversion: filming carriers coalescing into excitons

Authors: Enrico Perfetto, Takumi Fukuda, Xing Zhu... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How do free electrons and holes at semiconductor band edges convert into bound excitons? Specifically, what microscopic pathway bridges cold free carriers and their bound few-particle states, a step that has remained experimentally and theoretically elusive despite good understanding of individual carrier cooling.

Summary: The authors combine ultrafast momentum-resolved photoemission on monolayer WSe2 with first-principles excitonic-Bloch equations to directly observe free electrons and holes converting into bound excitons. They resolve a transient coexistence of free-carrier and excitonic bands with spectral-weight transfer, and show the process is a phonon-driven sequential cascade shaped by valley multiplicity, large-momentum phonon emission, and spin-flip transitions.

Key Results: Using ultrafast time- and momentum-resolved photoemission spectroscopy (trARPES) on monolayer WSe2, combined with first-principles excitonic-Bloch equations, the authors directly image the transient coexistence of free-carrier and excitonic bands and observe spectral-weight transfer between them. They establish a phononic origin of exciton formation via a sequential relaxation cascade to the lowest excitonic states with weakly populated intermediates, controlled by valley multiplicity, large-momentum phonon emission, and spin-flip processes. The abstract does not quote specific quantitative benchmarks (rates, timescales, or efficiencies).

Key Findings:

Free-carrier and excitonic bands transiently coexist during formation, with measurable spectral-weight transfer between them.
Exciton formation is phononic in origin and proceeds as a sequential relaxation cascade whose intermediate states remain weakly populated.
Formation efficiency is governed by valley multiplicity, large-momentum phonon emission, and spin-flip processes.

Technical Novelty: First direct momentum-resolved filming of the free-carrier-to-exciton conversion with sufficient energy resolution to resolve the coexistence of the two bands, coupled to first-principles excitonic-Bloch equations that assign the mechanism to phonon-mediated sequential relaxation with spin-flip and large-momentum channels — previous work characterized carrier cooling or exciton dynamics separately, not the conversion step itself.

What's New: Prior studies either tracked free-carrier cooling or characterized already-formed excitons; this work fills the missing middle by directly imaging the conversion in energy and momentum, and matches it to a first-principles microscopic theory that identifies the specific phonon channels responsible.

Extension Opportunities:

Apply the same trARPES + excitonic-Bloch-equations methodology to other TMDs (MoS2, MoSe2, WS2) and heterostructures to map how band alignment and interlayer coupling reshape the phonon-driven formation cascade.
Engineer phonon spectra via substrate choice, strain, or isotopic substitution to selectively enhance or suppress the large-momentum phonon channels identified as bottlenecks, and quantify the resulting exciton yield.
Build a predictive simulator coupling first-principles phonon spectra with excitonic-Bloch equations as a design tool for exciton-based devices (LEDs, photovoltaics, valleytronic switches), and validate against the trARPES observables reported here.

Replicability: No code or data availability is indicated in the abstract. Reproducing the experiment requires a high-resolution trARPES setup with monolayer WSe2 samples (ultrafast XUV/HHG source, hemispherical or momentum microscope analyzer, UHV). Theory reproduction requires DFT + many-body (GW-BSE) infrastructure and custom excitonic-Bloch-equation solvers — substantial HPC resources typical of first-principles condensed-matter groups.

Research Gaps:

Generality beyond monolayer WSe2 — how these mechanisms transfer to other TMDs, heterostructures, or bulk semiconductors is unknown.
Quantitative control knobs: which experimental levers (temperature, strain, dielectric environment, doping) can be tuned to accelerate or suppress each step of the cascade remain to be systematically explored.

🏥 HEALTHCARE

1. What to Remove, What to Preserve: Dual-Ambiguity Rectification for All-in-One Image Restoration

Authors: Cencen Liu, Wen Yin, Dongyang Zhang... Published: 2026-07-30 | Citations: 0 arXiv | PDF

Research Question: How can all-in-one image restoration models disentangle degradation cues from scene content to avoid the 'dual ambiguity' problem — semantic ambiguity in channel-wise modulation and spatial ambiguity in restoration responses — that causes content corruption and residual artifacts when handling heterogeneous degradations in a shared latent space?

Summary: DAR-Net addresses entanglement between degradation cues and scene content in all-in-one image restoration by introducing a structured degradation representation via simplex-constrained archetype mixtures, then rectifying both semantic (channel) and spatial ambiguities through prompt-based conditioning and orthogonal feature subspaces. It sets new SOTA on standard all-in-one benchmarks, with the largest gains (+0.34 dB PSNR) on the more challenging five-degradation setting.

Key Results: DAR-Net achieves best overall performance on standard all-in-one restoration benchmarks under both three-degradation and five-degradation settings, improving average PSNR over the strongest competitor by 0.14 dB and 0.34 dB respectively. Additionally shows superior performance on CDD-11 and WeatherBench benchmarks. Gains are larger in the harder five-degradation setting, suggesting the rectification approach scales with task heterogeneity.

Key Findings:

Explicitly separating degradation-aware from complementary content features via orthogonal subspace regularization reduces spatial artifacts
Simplex-constrained archetype mixtures produce a more discriminative degradation state than free latent embeddings, enabling better channel-wise conditioning
The dual-rectification approach yields larger PSNR gains as the number of concurrent degradations grows (0.14 dB → 0.34 dB from 3 to 5 degradations)

Technical Novelty: Three coupled contributions: (1) Degradation Archetype Representation (DAR) using simplex-constrained archetype mixture modeling to build a structured degradation state rather than a free-form latent embedding; (2) Semantic Ambiguity Rectification (SeAR) that generates degradation-aware prompts for channel-wise decoder conditioning; (3) Spatial Ambiguity Rectification (SpAR) that regularizes degradation-aware and complementary features toward orthogonal response subspaces. The explicit orthogonality constraint between 'what to remove' and 'what to preserve' features appears to be the core novel mechanism.

What's New: Prior all-in-one methods (e.g., PromptIR, AirNet) encode degradation in a shared latent space without explicit structure or separation from content. DAR-Net's contribution is framing the failure mode precisely as 'dual ambiguity' and imposing structured priors — simplex archetypes for the degradation state and orthogonality for the removal/preservation feature split — rather than relying on the model to disentangle implicitly.

Extension Opportunities:

Extend the archetype mixture modeling to video restoration where temporal consistency of degradation archetypes across frames could further reduce ambiguity
Apply the orthogonal subspace regularization (SpAR) idea to other conditional generation tasks like text-to-image where content and style entanglement is a known problem
Investigate whether the learned degradation archetypes transfer zero-shot to unseen degradation types, potentially enabling a universal restoration prior via archetype interpolation

Replicability: Abstract does not mention code release or dataset availability. Benchmarks used (three/five-degradation all-in-one, CDD-11, WeatherBench) are standard and publicly available. Compute is likely moderate — comparable to other restoration transformers (single-node multi-GPU training over days), though exact numbers are not disclosed in the abstract.

Research Gaps:

Behavior on unseen or compound degradations not represented in the archetype set is not clarified
Interpretability of learned archetypes — whether they correspond to human-meaningful degradation types — is not addressed in the abstract

🔥 GitHub Trending

1. Indhu27-pixel/Labour-Law-Violation-System

⭐ 33 stars | Python

AI-powered Labour Law Violation Detection System using Weak Supervision, BiLSTM, Graph Neural Networks (GNN), and Explainable AI to identify labour law violations.

artificial-intelligence deep-learning explainable-ai flask gnn graph-neural-networks

2. FTShare-Lab/agent-claim-network

⭐ 12 stars | Rust

面向终端的通用 AI 助手，让 Agent 之间共享可检索、可追溯且允许分歧的知识。让下一次判断，带着来源出发。

ai-agents knowledge-sharing llm mcp multi-agent-systems rust

3. Yijia-Xiao/FinanceHarness

⭐ 7 stars | Unknown

Landing soon...

finance harness llm

4. Commonwealth-Bank-of-Australia/PSAO-IJCAI-release

⭐ 5 stars | Python

ai ai-agent ai-agents ai-model benchmarking benchmarking-framework

5. fx-studio/vibe-coding-2026-essentials

⭐ 5 stars | Unknown

Sách miễn phí tiếng Việt về vibe coding cho Tester, PM, BA — người làm IT không viết code. 20 chương, 3 phụ lục, 54 hình. Học cách chỉ đạo AI coding và đánh giá kết quả: quy trình 5 giai đoạn, prompti

agentic-ai ai-coding business-analyst ebook free-book llm

6. Commonwealth-Bank-of-Australia/project-echo

⭐ 5 stars | Python

ai ai-agent ai-lab gemma4 llm transcription

7. Wu-beining/MKC-Loss

⭐ 3 stars | Python

Morpho-Kinematic Constraint (MKC) Loss for 2D/3D sparse curvilinear structure segmentation

image-segmentation loss-function medical-imaging pytorch topology

8. xiaoliu07-sudo/xiaoliu-pojia

⭐ 2 stars | Python

小liu破甲 - Codex CLI jailbreak prompt and one-click deploy tool. 基于CTF框架的Codex破甲提示词与一键部署工具

ai-security codex jailbreak llm prompt-engineering security-research

9. kissshhot/Awesome-OPD

⭐ 2 stars | Unknown

A curated paper list on On-Policy Distillation (OPD) & On-Policy Self-Distillation (OPSD) for LLMs — student rollouts + teacher feedback, covering white-box/black-box methods, OPD-RL hybrids, agentic

awesome-list knowledge-distillation large-language-models llm llm-agent on-policy-distillation

10. NerdBaba/finsentinel

⭐ 1 stars | Python

Real-time financial transaction monitoring: XGBoost fraud scoring (Stage 1) + hybrid RAG risk investigation via OpenCode Zen LLM (Stage 2). FastAPI, Qdrant, MLflow, Docker.

docker fastapi financial-services fintech fraud-detection llm

11. nadunchandrabahu/Multi-instance-learning-report

⭐ 1 stars | Unknown

Graduate survey report on Multi-Instance Learning: the theory of learning from bag-level labels, a taxonomy of MIL methods, and a critical review of recent object-detection research.

boosting classification computer-vision data-science deep-learning drug-discovery

12. yansirgood/researchX-skill-codex-local

⭐ 1 stars | Python

Behavior-preserving Codex port of phd-skills with embedded PaperHarness for auditable paper and baseline reproduction.

agent-skills codex machine-learning paper-reproduction paperharness research

13. bicerselimcan/bank-marketing-response-prediction

⭐ 1 stars | R

Customer response prediction using Logistic Regression, SVM, and Random Forest with 10-fold CV in R

classification cross-validation data-science machine-learning r random-forest

14. Mlishnii/fish-weight-prediction

⭐ 1 stars | Jupyter Notebook

Predicting fish weight from physical measurements using Decision Tree and Random Forest regression.

desiciontree fish jupyter-notebook machine-learning ml python

15. venkat4246/EduMentor-AI

⭐ 1 stars | TypeScript

AI-Powered Learning and Career Guidance Platform built with Streamlit.

artificial-intelligence educational-project gemini-ai learning-platform machine-learning python

Generated by Research Pulse on 2026-07-31 06:07